Research

Azton Wells’s research spans machine learning and astrophysics across two overlapping strands. The first develops high-resolution cosmological hydrodynamic simulations of the first galaxies. The Phoenix suite, built in Enzo, samples tens of thousands of primordial star-forming regions and resolves the heterogeneous chemical enrichment driven by Population III supernovae. StarNet, a 3D convolutional surrogate, accelerates Pop III star formation and feedback enough to be integrated directly into adaptive-mesh cosmological runs, enabling statistically significant galaxy samples at z ≳ 10 that remain physically consistent with higher-resolution reference simulations.

The second strand applies large language models to scientific practice. The AstroMLab series introduces astronomy-specific LLM benchmarks and releases AstroSage, a family of domain-specialized models (8B and 70B parameters) that match frontier general-purpose systems on astronomy Q&A at a fraction of the inference cost. Related work extends transformer architectures beyond text: LoRA adaptation of LLaMA-3.1-8B to predict galaxy redshifts from spectra, and encoder-only foundation models that map between simulation- and observation-derived galactic features with dynamic masking across modalities.

Adjacent work targets the infrastructure needed to deploy such models at scientific scale — HPC-resident retrieval-augmented generation over millions of papers, evaluation methodologies for AI research assistants, and credit assignment in multi-LLM agents — alongside empirical findings that context length alone degrades LLM performance independent of retrieval quality. Current directions push multi-modal astrophysical foundation models toward images, merger trees, and 3D fields, and toward unified agents that handle both quantitative modalities and natural-language reasoning.